Chasing the
real Jollibee

The timeline at a glance

The commissioned bee

The story starts on the morning of 24 August 2023, with three files arriving within three minutes of each other: a screen recording at 07:48, a Blender scene at 07:49 and an FBX at 07:51, all named rizky-jollibee. Today they sit on the DISK4 archive in a folder called 3d-rixky, oddly inside assets-3d-2022. The misfiling is given away by a JPEG exported a week later. Blender writes the path of the scene it rendered from into the JPEG comment, and that path reads …\budots-media\assets-3d-202308\rizky\jollibee\jollibee2.blend. The folder was originally "assets, 3D, August 2023", on a Windows PC, and "rizky" is almost certainly the name of the 3D artist who built the model.

The delivery is a tidy freelance package. The 1 minute 42 second video (1026 × 774, written by ffmpeg) is a screen capture of Blender in pose mode: the artist grabs the controls and bends the arms, spine and hands to show that the rig works. The Blender file (saved by Blender 3.2) holds the character as 13 separate meshes (Cube, Sphere, Torus, Plane…), 9,601 vertices in all, on a 65-bone Mixamo skeleton. The rig's bones are named mixamorig:Hips, mixamorig:Spine and so on, and it comes with a baked Mixamo animation. There's one material and one 4096 × 4096 texture, plus a reference image whose name follows Facebook's download pattern (310928400_651475766358131_…_n.png). The artist was working from a picture found on Facebook.

Within twenty minutes the model was opened in Blender 3.6 as jollibee.blend (created 08:09, saved until 28 August). The first proper render followed on 28 August at 15:38: Jollibee in a T-pose on a dark, starry set, glowing, arms out. It's clearly him. The peach face mask with the big eyes, the toque, the jacket, the white shirt with three buttons, the bow tie and the yellow belly are all there. It also shows the first of the problems that follow him for three years. The proportions are a little too human, the surfaces are flat, glossy plastic, and the face, the part people recognise first, is a painted mask on a smooth ball rather than a sculpted shape.

A 360° world called "Budots AR"

The second scene, jollibee2.blend (created 28 August 16:29, saved 2 September 11:47), is where ambition shows. It has 276 objects, particle "firebrand" and "light catcher" materials, glyph textures, and a completely new rig. Rigify generated a 614-bone armature called Jolibee from a 159-bone meta-rig: deformation, mechanism and control bones, IK poles for the arms and legs, an eye-target parent. That's ten times the Mixamo skeleton. On 2 September, between 09:13 and 11:39, it rendered a series of 5120 × 2560 equirectangular panoramas: Jollibee in a T-pose on a glowing floor, with the words ECOMMERCE and BUDOTS AR floating in the dark. It was a 360° showroom for a web-AR idea: walk around the bee on your phone.

The FBX with no body

On 15 September at 09:31 a new export appears: jollibee1.fbx, 7 MB, written by "Blender (stable FBX IO) 3.6.2". Opened today, it holds two armatures (the Rigify rig, now 952 bones once FBX adds its end bones, and the 198-bone meta-rig) and not a single mesh. Somebody exported the skeleton and forgot the bee. Twenty minutes later, at 09:52, a folder rizky-jollibee.fbm appears next to the original FBX, holding one file: color.jpg, 4096 × 4096. Programs built on Autodesk's FBX SDK (Unity, Maya and 3ds Max among them) unpack a model's embedded textures into a .fbm folder when they open it. So the fallback was to open the artist's original FBX in one of them.

That texture shows the second problem. The atlas is a flat colour-by-numbers sheet: red islands for the jacket and limbs, cream for the face, and several blocks of flat pink. Render the FBX by itself today, without the materials the artist had set up in Blender, and Jollibee wears a pink bib where his white shirt and yellow belly should be. The model only looks right inside the one scene it was made in.

Jollibee underwater

Two months later the approach flips. Instead of building a Jollibee, the idea is to capture one. The folder is called 3d-nerf/jollibee-underwater, and it holds 54 photos, G0043476.JPG to G0043529.JPG, from a GoPro HERO9 Black: 5184 × 3888, Protune on, 1/120 s at f/2.5, ISO around 1,900–1,950. The subject is a small Jollibee figurine on a round red base, standing on the tiled floor of a swimming pool. The camera circles it, close, the tiles bending in the wide lens, for 53 seconds.

The metadata holds a small mystery: two clocks that disagree. The camera's own clock says the first frame was taken on 6 November at 03:36:53, which would be a very late swim. But 20 of the 54 frames also carry a GPS stamp, and GPS time says 08:35:21 UTC on 5 November. The two clocks tick in step: from the first frame to frame 520 the camera clock advances 44 seconds and so does GPS time (to 08:36:05). So the GPS was live, and the camera's clock was simply 19 hours fast. The photos were taken at 4:35 in the afternoon of Sunday 5 November 2023, and the coordinates put the pool in Kuala Lumpur. That the GPS got a fix at all, on and off, means the camera was in shallow water just under the surface. The Philippines' most famous bee was captured in Malaysia.

The photos then waited six weeks. On 19 December 2023 at 17:04 two results appear, later moved to the drone-footage disk under 3d-pointclouds-ph2023/philippines. One is a 16.8 second fly-through (1152 × 832) with the Luma AI mark in its corner. The other is Jollibee_Underwater_1_point_cloud.ply, 22 MB and 825,446 points, each with a double-precision position and an RGB colour. Its header reads comment Created by Open3D.

Seen from the front, the NeRF is the most "real" Jollibee in this whole history, because it is a real object: moulded vinyl, real light, real water. Turn it a little and it smears. The pool's tiles bleed into streaks, the back of the figure is guesswork, and what you get is a cloud of coloured dots, not a surface. It can't be rigged or animated, and it can't be lit in a new scene. It's also a toy, not the mascot: squat, fused to its base, one hand stuck in a wave.

Teaching him to dance

In April 2024 the question changes from "does he look right?" to "can he move?". A new folder, 3d ph characters/3d-jollibee-rig, is created on Sunday 7 April at 16:16, and a minute later jollibee_rig.fbx lands in it. The header gives away its origin. The creator is "Mixamo, Inc.", and a path from Adobe's servers has leaked into the file: /home/app/mixamo-mini/tmp/outputexport.fbx. It's the same bee design as the 2023 model, now exported as 13 flat-coloured parts (18,761 vertices, Maya-style lambert materials, no textures at all) and auto-rigged on a 57-bone Mixamo skeleton. The first thing Mixamo was asked to do with him, at 16:21, was Rifle Aiming Idle.fbx. Nobody said testing was dignified.

The next afternoon, 8 April at 15:51, comes 2024-budotsgames-jollibee-vert.mov: 20 seconds, vertical 1080 × 1920, 24 fps, encoded with "H.264 NVIDIA", the GPU encoder used by screen recorders on a GeForce PC. It shows Jollibee walking around a grey test level under a blue sky, seen from behind like a game character. It's the first time he's a player. On the same afternoon two brand images are saved as references: the "Jollibee Easter Wonderland" poster (31 March 2024, Robinsons Iloilo) with the whole gang, and a 201 × 251 thumbnail of the "Kayo pa rin ang BEE-da!" campaign.

A TikTok dance, through DeepMotion

On 11 April the folder jeliaandres-video-deepmotion appears. Inside is a 15 second vertical clip of a Jollibee dance by the actress and creator Jelai Andres (576 × 1024, and the vid: tag in its metadata is a TikTok video ID). Two FBX files follow at 15:59. One is named …-no-watermark_…(includeTPose).fbx, which matches the export options of DeepMotion's Animate 3D, the service that turns an ordinary video of a person into skeletal motion. Both FBX files name Maya as the program that wrote them, on DeepMotion's side.

That evening at 19:46 a Blender 4.0.1 scene is started: blender-jollibee-andres-tiktok.blend, 231 objects. The dance data (Armature|Jelai_Andres_…|default) drives Jollibee's Mixamo skeleton. The scene also holds an Avaturn full-body avatar (avaturn_body, avaturn_glasses, avaturn_shoes), a chimp, a diorama and cloud kit, and an "aircraft workshop" HDR for lighting. By 09:31 the next morning the 300-frame render (12.5 s, 1080 × 1920, 24 fps) is done, and at 11:28 and 11:31 DaVinci Resolve Studio writes two edits, 2024-jollibee-rig.mp4 and 2024-jollibee-rig2.mp4. In them Jollibee dances in purple light, with the original dancer in a picture-in-picture in the corner. At 15:43 a 1920 × 1920 poster follows: budots-dance-academmy.png (sic), Jollibee and the avatar posed in front of the words "Budots Dance Academy".

By April 2024 motion was solved: any dance on TikTok could be on Jollibee by the next morning. What wasn't solved was Jollibee.

The dance works, and it looks right from a distance. Up close, it's the 2023 model: stiff shell parts that intersect when the hips roll, a flat decal face, arms that bend like tubes. The rig was never the weak point. The body was.

The rig comes back, dancing

The same rig turned up once more on 28 September 2026: Jollibee_dancing_Twerk.usdz, 1.7 MB, holding one scene.usdc written at 01:14. It's a Sketchfab export of this model (18 meshes, about 20,500 vertices, the Maya lambert materials) on the same 57-joint Mixamo skeleton, renamed n33 to n89 by the conversion, with a Mixamo twerk: 457 samples, 15.2 seconds at 30 fps. Measured from the file, the hips bounce at 106 beats a minute and drop by up to 38% of their height. It also has faults. The animation is bound to each mesh rather than to the skeleton, so strict USD tools don't play it. And the neck turns the head up to 103° against the chest; at the extremes the face mask and the head shell come apart. The motion itself is good, so the Fiesta now uses it: retargeted live onto the 2026 model, with the neck held to a neck's range (see below).

The brand-kit night

The next folder was built in a single session between midnight and half past two on 20 December 2024. It starts with references. At 00:01 a press photo of the costume mascot with Jollibee Foods' CEO, Ernesto Tanmantiong, and at 00:03 a YouTube thumbnail titled "Bringing Joy Online". Then, from 00:37 to 00:41, six video files: two clips of the brand's own CGI film for its delivery app (Jollibee in a glossy city of hexagon platforms, a giant Chickenjoy bucket, the caption "Introducing the all-new Jollibee App"), each in three broadcast sizes: standard definition (720 × 486 and 720 × 576), HD and 4K. That's the kind of deliverable set a production agency hands a client. The 4K file alone is 205 MB.

This is the benchmark: what the real Jollibee looks like in 3D when the brand makes him itself. At 01:26 a DaVinci Resolve Studio re-cut (jollibee2_1280x720.mp4) suggests the clip was studied frame by frame. At 01:59 the official vector artwork is saved: Jollibee-head.svg, JollibeeText.svg, Jollibee-Logo.svg, and at 02:30 the Illustrator original, Jollibee-Logo.ai.

At 02:31 the night's own attempt appears: jollibee.usdz, 12.8 MB. Its model.usda says creator = "Three.js USDZExporter", and its scene is anchored to a horizontal plane, which is what iPhone AR Quick Look needs to put an object on a table. The SVG head and the wordmark were traced into shapes and extruded, 38 meshes in all (Jollibee_head_140, Shape___143…), so that you could stand the logo on your desk in augmented reality.

It's the only attempt that is on-model by definition, because it is the logo. It's also the least like a character. A badge has no back and no body, and it can't do anything. The small ._ companion files dated 1 January 2025 at 21:54 show when the folder was copied to the DISK8 archive: on New Year's Day.

Let the video models do it

By February 2025 image-to-video models had become good enough to try skipping 3D entirely. On 12 February at 19:33 a folder named 2025-02-kaishou-kling1_6 (the Kuaishou typo is in the original) gets four reference pictures in two minutes: Jollibee-shop.jpg, a Jollibee storefront with the full-size mascot statue waving outside; the 2011 logo; a single piece of fried chicken; and a YouTube-style thumbnail of four schoolboys with a Jollibee bucket.

Then the generations, all 5.1 seconds, 1280 × 720, carrying the "Kling AI 1.6" watermark. Kling names its downloads after the start of the prompt, so the prompts survive only as those fragments:

• Standard_Mode_16x9_add_splash_of_yellow_and_red_par… · 20:06
• Standard_Mode_16x9_add_filipino_children_jumping_ha… · 20:35 (twice)
• Standard_Mode_16x9_add_filipino_kids__add_yellow_an… · 20:37 (twice)

After midnight (13 February, 01:52–01:54) food references follow: a fried-chicken commercial still and two photos of fried chicken. On 17 February at 19:11 comes a clip made with a different model: jollibee-minion.mp4, 5.6 seconds, 1088 × 720, carrying the MiniMax / Hailuo AI watermark. Jollibee in costume high-fives a Minion across a fast-food counter. The same week, a sibling folder, ph whangod 2, fills with clips of the tattoo artist Whang-Od in Tokyo, on a longboard, as a pilot. Jollibee was one of a series of Philippine icons run through the new video models.

These are, frame for frame, the most photographic Jollibees in this history. That's because the models animate a photograph of the real statue and the real costume. They also show the limits of the method. Each clip lasts five seconds. The face drifts when he turns (the statue's eyes slide, the toque changes shape), and there is no way to ask for the same Jollibee twice. What you get is footage, not an asset: nothing you can pose, light, put in a game or hand to a client as a character.

The go-kart morning

On Tuesday 27 May 2025 an idea goes from picture to 3D model in about an hour. The folder is 3d-model-turbosquid2hunyuan/jollibee, named after a pipeline for turning stock models into AI-generated meshes, though here the input is AI pictures. The chatgpt-draft subfolder holds the source images, each carrying a C2PA provenance manifest: claim generator "ChatGPT", software agent "GPT-4o", digital source type trainedAlgorithmicMedia, signed with Truepic's signing tool as ChatGPT does. One is a 1024 × 1536 cartoon of Jollibee racing a go-kart numbered 1 round a track. The other is a 1536 × 1024 model sheet: the same Jollibee as a soft vinyl toy in his kart, from five angles.

Single views are saved one by one: front.png at 11:04 (580 × 598), back.png at 11:21, left.png at 11:23, right-isometric-jollibee.png at 11:24. At 11:29, jollbee-gokart-hunyuan25.glb: 45.8 MB, from Tencent's Hunyuan3D 2.5 multi-view image-to-3D model (the "hunyuan25" is in the filename). It's one mesh with 307,983 vertices and exactly 499,999 triangles, the generator's half-million-face ceiling, and four 4096 × 4096 PBR maps named texture_pbr_v128. A turntable video follows at 11:30, an FBX at 11:43 (68 MB, Blender 4.0) and one last angle at 12:06.

This was the first time a 3D Jollibee looked like the modern Jollibee: round, soft, toy-like, the face mask the right peach and the eyes big and friendly. In an hour, from five pictures. But it can't be taken apart. Bee, steering wheel, gloves and kart are one continuous skin; the hands are melted into the wheel. There's no skeleton, and the mesh is thirty times heavier than a phone game can afford. Look behind him and the textures turn to mush where the generator had to invent what the pictures didn't show.

The next evening, 28 May at 22:11, a Gemini image is saved twice into a new folder, 3d-jollibeee (three e's): 2048 × 2048, with IPTC metadata "Made with Google AI" and the digital source type trainedAlgorithmicMedia. It shows a faceted, low-poly-looking Jollibee standing in a neutral pose, twice. It's a picture of a 3D model rather than a 3D model.

The go-kart picture then went round the internet. On 24 October 2025 it turns up as jollibee-gokart.jpg, 559 × 422 and 41,125 bytes, in the Q4 2025 isometric-art folder. A byte-identical copy sits next to it as 511289820_4304268413231239_4906238356963253133_n.jpg, which is Facebook's naming pattern. The render had been posted to Facebook and downloaded back. That small, recompressed copy is the version that went into the 2026 game pitch.

A full-body likeness

On the evening of 25 November 2025, in the Q4 isometric project folder, two references are saved: at 19:38, Jollibee.jpg, the official full-body illustration (hands clasped, one foot forward); at 19:42, jollibee-mascot-new-restautant-tst-hk.jpg, a photo of the costume mascot at a new restaurant in Tsim Sha Tsui, Hong Kong. At 20:51 jollibee.glb appears in 3d-jollibeee, 41 MB, last saved at 22:59 by Blender 5.0's glTF exporter.

Its texture names, texture_pbr_20250901 with normal and metallic-roughness maps, follow the same scheme as the May go-kart's texture_pbr_v128, now with a 1 September 2025 version stamp: a newer release of the same kind of image-to-3D generator. It's 266,334 vertices and 453,311 triangles in a single mesh, with two sets of vertex colours on top of the textures.

This is the closest any generated model came. It has the illustration's pose, the jacket over the striped belly, the bow tie, the gloves, and the glossy finish of a collectible vinyl figure. It also shows how generators read pictures: the illustration shades the white toque with pale blue, and the model faithfully made the hat blue. And like every generated mesh, it's one welded surface. The clasped hands are fused to each other and to the belly, and the arms are fused to the body. To animate it you'd have to cut it apart and rebuild everything the cut exposes, which is the job the final attempt ended up doing on a different model.

Jollibee Quest, the pitch

In August 2026 the go-kart Jollibee becomes a game pitch, made in Hilo, an AI canvas app whose project folder .hilo keeps a workflow graph, a thumbnail cache and even its own memory notes. The folder budots-games-jollibeecart is created on 10 August. The next morning two pictures are pasted in (their filenames are the Unix time of the paste): at 11:16 the Facebook-sized go-kart Jollibee, and at 11:17 an isometric diorama of Cebu City from the studio's own "GenAI by Budots Media Philippines" series, dated 29 November 2025.

The graph's edges record what was made from what. At 11:25 the two pastes became racing-isometric-track.mp4: 7.3 seconds, 1440 × 2560, Jollibee driving through the toy city past a red-roofed kiosk and the old stone church. At 13:14 the two pastes became a landscape menu, co-op-menu-confirmation.png (2560 × 1440), titled "CO-OP QUEST", with Player 1 Jollibee and Player 2 a blue wolf. At 13:24 came a clip called budots-media.mp4. Then a correction, preserved in Hilo's memory file at 13:45: "Use Jollibee Quest as the title/header text. Not 'Coop Quest' or other variants." A second note says "Use 9:16 aspect ratio (portrait/vertical) as the default." The portrait "JOLLIBEE QUEST" menu is dated 13:44, and at 13:51 an animated version, animated-game-ui.mp4 (10.1 s). The canvas remembers the last video settings: model MiniMax-H3, reference-image mode, 10 seconds, 9:16, 2K, with generated audio.

On 30 August everything was copied to a folder called vid-hy3, with the landscape menu renamed jollibee-karting-menu-confirmation.png (byte-identical). On 25 September at 03:06, the concept went up on budotsmedia.com as "Jollibee Go-Kart in Cebu, a game concept".

It's a good pitch, and it shows the gap exactly. In the trailers Jollibee is a sticker: a picture of a toy, slightly different every time. To turn "Jollibee Quest" into something you can play, you need a Jollibee you can control. Everything so far had failed that test.

The sculpt that became a game

On 25 September 2026 at 22:02 a 23 MB file arrives: Jollibee_welcome.obj. Its first line reads # File exported by ZBrush version 2021, then #Vertex Count 282716, #Face Count 565444, #Auto scale x=0.013333. It's a digital sculpture of Jollibee in his welcome pose, one arm out, the other hand on his belly. It has no colours, no UV map, no materials and no skeleton, just one closed surface, the kind of file made for printing or presentation. It was supplied for the project.

Its advantage over everything before it is the one thing that can't be generated reliably: a sculptor's understanding of the character. The face mask is a real raised edge, not paint. The eyes are carved hollows, the grooves between jacket, shirt, belly stripes and gloves are cut into the surface, and the proportions are the mascot's, not a person's or a toy's.

Painting a sculpt that has no paint

Over the next two nights a pipeline of small Python and Blender scripts (tools/jollibee-model/ in the game's code) turned the sculpt into a game character. They used those carved grooves as the paint guide.

First the mesh was exported as arrays. Then every vertex was scored for how deeply it sits in a groove, in two ways: the curvature at that point, and a cavity render. Blender rendered the sculpt from 24 orthographic directions, twice each, once with every face coloured by its own index number and once with cavity shading, and projected the dark lines back onto the faces. The belly stripes are shallow steps that only the cavity render finds. Cut the surface along the grooves and it falls into regions: face mask, eyes, toque, band, jacket, shirt, bow tie, buttons, gloves, stripes, shoes, wings.

Each region was then painted after the published design: red head, peach face, rosy cheeks, a red nose, huge black eyes with white glints, white toque, red jacket, white shirt, black tie and buttons, white gloves, red and yellow stripes, yellow shoes and light-blue wings. The seams between regions were painted as soft, darker shades of the colours on either side rather than black lines. The build decimated the model to about 34,600 triangles, gave the head more than twice the texture detail of the body, baked the painted high-resolution sculpt onto the light mesh as a colour map and a normal map (4096² reduced to 2048²), put the wings on their own transparent material, built a 10-bone skeleton with weights per region, and exported a Draco-compressed GLB of 1,078,416 bytes.

The arm that was welded to his belly

The sculpt had the problem every generated model had: his right arm hugged his side and the glove rested on his belly, welded where they touched. Version 1.0 of the game (26 September) shipped with it. Version 1.2, the same day, cut the arm free along the weld everywhere except the shoulder, capped both openings, painted the caps with the colours at their rims and moved the shoulder joints to the measured arm roots. Now he could lift both arms in a V to cheer, or spread them like wings when he jumped. Version 1.3 (27 September, git commit at 17:23) fixed the last of it. The hand no longer sinks into the belly while running, and the side of the jacket that had been hidden under the arm was rebuilt by mirroring the other side, so it shows as a smooth red jacket instead of streaks.

The result is Jollibee Island Run, an endless runner for kids from the CCLEX in Cebu across 22 places in the Philippines. Jollibee runs, jumps, flutters, gets dizzy, cheers and waves, animated live in the browser from ten bones, on an ordinary phone.

By the numbers

Every figure below was measured for this article by opening the original file in Blender 5.2 or reading its header, not taken from memory.

Two patterns stand out. The generated models are all enormous and all skeleton-free: 40 to 46 MB of single, welded surface. And the winning model is the smallest file in the table, about 40 times lighter than the generated ones, yet it has the best likeness, because its detail is in a 2048² baked texture and normal map instead of in half a million triangles.

What it took

Likeness is sculpted, not generated

The image-to-3D models of 2025 got astonishingly close to the silhouette in minutes. What they couldn't give was structure: they read Jollibee as a picture, so they welded what touches in the picture (hands to belly, bee to kart) and took pale-blue shading for a blue hat. The sculpt knew where the jacket ends and the shirt begins, because a person cut that line.

A character is a picture plus a skeleton plus a budget

Most of the attempts produce something that looks like Jollibee in a single frame. Only the rigged ones could move (the 2023 model, its 2024 re-rig, and the 2026 model), and only the last one looks right while moving and runs on a phone. The difference is unglamorous: topology, a rig, clean separations where limbs meet the body, and a file small enough to download before a child loses patience.

Motion was never the hard part

By April 2024, a TikTok dance could be on a 3D Jollibee by the next morning (DeepMotion, Mixamo, Blender). The hard part was a body worth animating. The dances are coming back now, on the new model: see the Fiesta below.

The files keep a diary

Nobody wrote this history down, but the files remembered it. A JPEG comment kept the original folder name and exposed a misfiled year. An FBX header leaked a path from Mixamo's servers. Kling filenames kept the prompts. C2PA and IPTC say which AI made which picture, and a Facebook filename shows where an image had been. Two clocks in one GoPro disagree by 19 hours, and GPS settles which one is right. It's worth naming files well and keeping the .fbm folders: they turned out to be the only record of three years' work.

What's next

With a Jollibee who can move, the older ideas come back. The dance experiments of 2024 and the go-kart pitch of 2025–26 are now running on the new model, along with twenty small games built to find out which looks, cameras and kinds of play suit him best: